New AI tool keeps hackers out but might lock users in too

Earlier this year a lab quietly rolled out a chatbot that could crack passwords as easily as it answered trivia. Researchers fretted that criminals might swap it for their old toolkits and launch a wave of break-ins. After months of testing, the same lab is now shipping a stripped-down edition that trades raw power for peace of mind.The cut-down model borrows simpler tricks from earlier software so it can still answer everyday questions without tripping alarm bells. Guardrails now block any request that smells like a network probe or a biology experiment—two fields where rogue AIs have historically caused trouble. The idea is to keep honest users honest and, hopefully, keep dishonest users from getting ideas. Yet the trade-off worries some security teams. If the guardrails are too tight, defenders might find it just as hard to test their own systems or spot new tricks used by real attackers. The new tool is safer by design, but safety locks can also slow down the very people paid to keep networks safe.